Efficient and distributable methods for solving the multiagent plan coordination problem
Multiagent and Grid Systems - Planning in multiagent systems
| Hi-index | 0.00 |
The Multiagent Plan Coordination Problem arises whenever multiple agents plan to achieve their individual goals independently, but might mutually benefit by coordinating their plans to avoid working at cross purposes or duplicating effort. Although variations of this problem have been studied in the literature, there is as yet no agreement over a general characterization of the problem. In this dissertation, I describe a general framework that extends the partial-order, causal-link plan representation to the multiagent case, and that treats coordination as a form of iterative repair of plan flaws between agents. I show, analytically and empirically, that this algorithmic formulation can scale to the multiagent case better than can a straightforward application of the most advanced existing coordination techniques, highlighting fundamental differences between my algorithmic framework and these earlier approaches. I then examine whether and how the Multiagent Plan Coordination Problem can be cast as a Distributed Constraint Optimization Problem (DCOP). I use ADOPT, a state-of-the-art system that can solve DCOPs in an asynchronous, parallel manner using local communication between individual computational agents. I extend the ADOPT framework to take advantage of problem structure, in an effort to improve its performance on representative Multiagent Plan Coordination Problems. Although the performance gains from using ADOPT were negligible, I show how the same problem structure can be used to enable my algorithm to operate more efficiently. In the interest of scaling my approaches to larger problems, I build on previous work in the area of multiagent coordination [8] to make use of abstraction techniques that create hierarchical planning structures that are decomposed into primitive actions. By exploiting past work on developing summary information at abstract levels of agent plan hierarchies, I then extend my coordination algorithm to reason about agent interactions at abstract levels of the agents' plan hierarchies, and empirically establish that this technique greatly reduces the amount of computation required to derive coordinated planning solutions for the agents. I conclude with a discussion of possible ways of extending my work to handle richer and alternative plan coordination problems.